RSS Feed
30.5.14 | 
PhD Student Position: Sparse Representation and Learning in Pattern Recognition
at Computer Science Department, Laval University, Canada
Keywords:
Sparse Representation, Structured Sparsity, Sparse Subspace Learning,
Visual Recognition, Feature Selection, Sparse Coding, Sparsity Induced
Similarity,
Background:
Sparse representation and learning have been extensively used recently
in machine learning, computer vision, pattern recognition, etc.
Generally speaking, sparse representation and learning aim to find the
sparsest linear combination of basis functions from a complete
dictionary. A rational behind this lies in the fact that there is a
sparse connectivity between nodes in human brain.
In
many signal processing applications (video, image processing, speech
recognition, etc.) the data sets are usually high dimensional and very
large. In this context, sparse representation and learning have shown to
be promising techniques for addressing them.
Recently,
many important theoretical results enriched this area as for instance
[1]: (i) the sparsest representation in a general dictionary is unique
and can be found by using L1 minimisation [2]; (ii) the sparse
representation can be covered by solving the convex programming, if the
dictionary has a restricted isometry property [3,4]. Thanks to these
important results and their corollaries, sparse representation and
learning have extensively been used in many areas including signal
processing and applications as speech recognition, machine learning,
computer vision, digital multimedia, robotics, etc. A complete review on
sparse representation and learning from both theory and applications
sides appeared recently [1].
Goal and Objectives:
The goal of this PhD research is to strive to address the following issues:
- Is the sparsity assumption always supported by the data? Nowadays, compressive sensing has become one of the standard techniques of object recognition. If the sparsity is however not supported by the data, it is not guarantee to recover the exact signal and therefore sparse approximations may not deliver the robustness or performance desired [5]. In this case what sort of acceptable (in terms of computation load) robust method can be?
- When the Sparse Representation is Relevant? It is important here to perform an in-depth analysis of sparse representation in pattern recognition and see empirically if this sparse representation improves recognition performance compared to non-sparse representations [6]. To this end, it would be important to take into consideration the way to extract features and to refine them as well as the computational load induced by the all process of sparse representation.
- Sparse Representation or Collaborative Representation, which one is the best? Similarly to Zhang’s work [7], it would be appropriate to see if the use of all training samples to collaboratively represent a query sample is much more crucial to sparse representation based classification (SRC). Taking into consideration the fact that the collaborative representation based classification (CRC) plays a more important role than L1-regularization as shown by Zhang et al. [7], it would be opportune to see what new instantiations of CRC (with less computational load than usual SRC) can be proposed.
- Is sparse representation and learning usefulness in the context of video-based action modeling and recognition? Are ideas from this application fairly general and applicable to other recognition problems? One should here explore the usefulness of sparse representation and learning in the context of video classification, looking particularly at the problem of recognizing human actions-both physical actions and facial expressions [8]. This can be achieved by constructing an overcomplete dictionary using a set of spatio-temporal descriptors (extracted from the video sequences) in such a way that each of these descriptors is represented by some linear combination of a small number of dictionary elements. By doing so, one can achieve a more compact and richer representation than classical methods using clustering and vector quantization. It is also important to see which representation (sparse vs collaborative) is the more convenient for human-action recognition. Experiments and validation of generalization to other recognition problems can be done on several data sets containing various physical actions, facial expressions and object recognition.
Job Description:
The
PhD candidate will focus on signal processing and machine learning. In
this context, she will first acquire expertise in different topics such
as clustering and classification, Bayesian and generative modelling,
signal separation, parameter and state estimation, time series and
space state methods, compression and coding. Then, the PhD candidate is
expected to contribute to the advancement of the literature on sparse
representation and learning along many different lines: methodological,
theoretical, algorithmic and experimental.
Profile:
The
applicant must have a Master of Science in Computer Science or Computer
Engineering, Statistics, or related fields, possibly with background in
Signal Processing and optimization. Good written and oral communication
skills in English are required.
Application:
The
application should include a brief description of research interests
and past experience, a CV, degrees and grades, a copy of Master thesis
(or a draft thereof), motivation letter (short but pertinent to this
call), relevant publications (if any), and other relevant documents.
Candidates are encouraged to provide letter(s) of recommendation and
contact information to reference persons. Please send your application
to chaib@ift.ulaval.ca. The deadline for the application is June15th, 2014, but we encourage the applicants to contact as soon as possible:
- Supervisors: Pr. Brahim Chaib-draa
- Place: Department Computer Science and software Engineering, Laval University, Québec
Working Environment:
- The PhD candidate will work at Damas lab part of the Department Computer Science and software Engineering in collaboration with researchers from REPARTI center of
which Chaib-draa is affiliated (http://reparti.gel.ulaval.ca/
). REPARTI provides several advantages, including a human infrastructure to help with some management aspects for research, and access to a wide network of other researchers in areas closely related to the topic of this PhD proposal.
Benefits:
- Duration: 36 months – starting date: September 2014, 1st
- Salary: 19 000$/Year + 3 000$/year (from University)
References:
- Cheng, H.; Liu, Z.; Yang, L.; and Chen, X. Sparse Representation and Learning in Visual Recognition: Theory and Applications, Signal Processing, 93, 2013.
- Donoho, D. and Elad, M. Optimally Sparse Representation in General (non-orthogonal) Dictionaries via L1 minimization. Proc. Of the National Academy of Sciences, 100(5), 2003.
- Candes, E.J.; Romberg, J. K. and Tao, T. Stable Signal Recovery from Incomplete and Inaccurate Measurements. Communications on Pure and Applied Mathematics, 59(8), 2006.
- Candes, E. J. and Tao, T. Near optimal Signal Recovery from random Projection: Universal Encoding Strategies? IEEE Transaction on Information Theory, 52(12), 2006.
- Shi, Q,; Erikson, A.,; Hengel, A. and Shen, C. Is Face Recognition really a Compressive Sensing Problem? In Proc. of CVPR’11, 2011.
- Rigamonti, R.; Brown, M. A. and Lepetit, V. Are Sparse Representations Really Relevant for Image Classification? In Proc. of CVPR’11, 2011.
- Zhang, L.; Yang, M.; and Feng X. Sparse Representation or Collaborative Representations: which Helps Face Recognition? IEEE Int. Conf on Computer Vision, 2011.
- Guha, T. and Ward, R. K. Learning Sparse Representations for Human Action Recognition, IEEE Transaction on PAMI, 34(8), 2012.
